monthly-model-review v1

1 Scorecard - last month’s calls vs what the tape did

No prior monthly review exists. This is v1, so there are no per-name calls to grade and no direction to check against the revision slopes. The scorecard begins next month, when the September edition grades these eight calls against the realized slope, stance and expectations moves.

One forward note for that grading: two of the eight calls this month are explicitly premise-vs-condition splits that the next scorecard must not collapse into a single verdict - 6981 (the revision-cycle premise is spent on the near year even though its falsification condition is intact) and CBRS (the OpenAI-dependence test is technically satisfied but mis-specified). Both are flagged in their name blocks below; both are calibration-ledger candidates for the analyst if they resolve the wrong way, never auto-logged.

2 Per-subsector - what changed this month

Read this section as a two-day delta, not a month. The implications layer was seeded in the middle of August, so the only _history snapshots that bracket anything are 2026-08-20 and the current 2026-08-22 file. Seven of twelve themes show a direction flip across those two days, which is far too fast to be a real regime rotation - most of it is the implications layer’s own synthesis maturing as more of the print season was ingested, not two days of market movement. The stance-line rewrites below are the signal worth keeping; the direction flips are noise until the window is a real month.

• SPE / WFE (positive to mixed). The load-bearing rewrite of the month. Old vintage: “every SPE vendor raised this window on TSMC USD64bn capex and memory restarts; risk has migrated from estimates to a multiple that de-rates into the raises.” New vintage: “prints are uniformly strong but the trade split twice - slope now favours test, packaging and content, while price embeds 5.7-8.4 percent terminal growth everywhere except KLA and ASMPT.” This is the whole edition in one line: the raises are real and universal, the multiple is the risk, and within the group the revision slope now favours test (6857) and process-control over broad WFE. Four of my eight names live here.
• Memory / HBM (positive to mixed). Old: “capacity sold out through 2026 and pricing still climbing… own conversion winners over HBM share stories.” New: “prices still rise but price no longer converts to margin and take-or-pay locks volume not price; the tape already prices a violent fade, so the edge is test intensity, not memory ASP.” The MU brief confirms this from the primary side - management itself conceded price-to-margin conversion is decaying.
• AI / Compute (positive to mixed). Old: “the binding constraint has moved off the accelerator into the bill of materials.” New: “not a fade but a rotation - the revision impulse has left large-cap compute for test, process-control and packaging, while price still embeds terminal growth far above the 2.5 percent anchor.” The AMD brief is the case study: MRVL +13 percent against AMD -12 percent on a flat AI tape - the market paid the custom-ASIC name and marked down the merchant.
• Passives / MLCC (positive to mixed). New stance: “the shortage is not in FY26 numbers - EPS FY26 revision slope is 0.0 percent at all three Japanese majors - it is in out-years the multiple already capitalizes at 2-3x consensus terminal growth.” The 6981 brief independently found the identical thing: FY27 EBIT consensus sits above management’s own guide, so the near-year revision premise is spent.
• Advanced Packaging (unclear to positive). The one genuine information-gain flip rather than a maturation artifact - the theme went from “quiet window, fewer than five triaged items” to a real read as the packaging prints landed. Not a held name in my eight, noted for completeness.
• Foundry, China Semis, Semi Materials (positive, held), Consumer Electronics and DC Power / Electrification (mixed, held) showed no material stance movement across the two-day window.

3 Per held name - the model calls

Grouped long first, then short. Each block is the deep-tier read from the name’s fresh brief; the full derivations, honesty flags and recompute recipes live in System\Initiate\{T}\research_brief_v1.md.

Longs

KLAC - EPS up, multiple down (terminal unreadable)

• EPS chain - revise UP. Revenue up: post-print consensus moved FY27 revenue +4.4 percent, FY27 EBIT +7.8 percent, FY27 EPS +5.3 percent, all of it after the 28 July print, on a Q1 FY27 guide of USD4.0bn, +24.6 percent YoY [evidence:KLAC:klac-ev-0480] [slope:KLAC:EPS FY26 (USD)] [expectations:KLAC]. Margins up but contested: the guide carries non-GAAP GM to ~62.5 percent, but the CFO ruled out repricing booked orders, so the raise depends on new-product pricing only, and a Q4 AR build (+27.6 percent YoY against +11.7 percent revenue) sits under it [evidence:KLAC:klac-ev-0482]. Shares/BVPS inert.
• Multiple chain - revise DOWN; terminal unreadable. Sentiment down: forward P/E de-rated -18.4 percent (43.8x to 35.8x) on a price move of only -7.4 percent - denominator, not price. Peer co-movement down: KLA is the persistent laggard of its own WFE complex (rebased 96.4 vs peer basket 108.0, -11.6pp; divergence fired 2026-08-05). Terminal leg cannot be read - the valuation engine’s reverse-DCF is contaminated by the known 10:1 yahoo-vs-CIQ split mismatch and is suppressed.
• Thesis anchor. Rests on the premium-durability claim [klac-th-001] and its paired durability claim [klac-th-003], the only two with stable meaning after the 2026-08-17 thesis rebuild. Load-bearing flag: the ledger’s bears_on tags predate that rebuild, so thesis_pulse counts for the premium-durability, service-annuity and China claims are contaminated - this review does not read the KLAC pulse counts at face value, per the brief’s own warning. The share-ceiling claim [klac-th-006] has zero evidence ever.
• Highest-signal datapoint. UBS’s flat CY26 WFE share (~7.7 percent) computed off management’s own raised USD150bn WFE guide reframes the beat-and-raise as CY26 market beta, not KLA share gain [expectations:KLAC].
• Biggest gap. Re-tag the evidence ledger to the current claim set and backfill the share-ceiling claim - /thesis backfill KLAC.

AMAT - EPS up, multiple down

• EPS chain - revise UP. Revenue up: the Q4 guide (USD10.25bn) landed 6-9 percent above the USD9.62bn street, the third consecutive raise, and China flipped from cap-risk to a guided CY26/CY27 growth driver [evidence:AMAT:amat-ev-0585] [news:amat-jul26-guidance]. FY27 EPS revised +8.5 percent in-window [slope:AMAT:EPS FY28 (USD)]. Margins confirm on the floor (thirteenth straight quarter of YoY GM expansion, record 50.3 percent GAAP GM) but crack at the top: Q4 GM guided flat and the peer-GM gap inverted ~150bps with no CFO mechanism [evidence:AMAT:amat-ev-0587] [evidence:AMAT:amat-ev-0589].
• Multiple chain - revise DOWN. Forward P/E de-rated 35.9x to 29.2x (-18.6 percent) on a -4.1 percent price move - a de-rate on an up-revision, and Fwd PEG at 0.98 sits in the 0.8th percentile of the trailing year.
• Thesis anchor. The month’s sharpest event is the event-path decomposition: the print popped +1.0 percent on release, faded through the call to -4.65 percent, and closed t+1 at -5.1 percent with excess vs SOXX of -5.06pp - idiosyncratic. This is the first observed instance this cycle of the de-rate-risk claim [amat-th-004] materializing on good news. Adjudicated AMBIGUOUS-leaning-live, with a claim-specification mismatch flagged: the risk arrived through the wrong door - the claim’s trigger is a downward revision, but estimates rose. The leading-edge-share claim [amat-th-002] stays AMBIGUOUS - all 25 in-window entries are self-reported, no independent WFE-share series exists.
• Highest-signal datapoint. The event-path ladder - the call, not the print, destroyed the reaction [evidence:AMAT:amat-ev-0585].
• Biggest gap. No independent WFE-share series, the deciding observable for the leading-edge claim - /deepen-corpus AMAT --tier broker-share then /thesis backfill AMAT.

6857 Advantest - EPS up (high quality), multiple down on applied / hold on realized

• EPS chain - revise UP, the highest-quality raise in the group. Revenue up on a company re-forecast, not a nudge: FY3/27 sales guide +20.7 percent to JPY1,714bn, OI +34.8 percent to JPY846bn, and the CY26 SoC-tester TAM raised to USD10.5-11.5bn from USD8.7-9.5bn [evidence:6857:6857-ev-0220] [evidence:6857:6857-ev-0221] [slope:6857:EPS FY27 (JPY)] [expectations:6857]. Consensus revised FY27 EPS +21.3 percent and has now passed the top-line guide. Unlike 6981, FX is secondary here. Margin haircut: forward EBIT margin ~49 percent is near double the 25 percent through-cycle model - peak-cycle risk - and growth more than halves from FY27 to FY29 [evidence:6857:6857-ev-0223].
• Multiple chain - revise DOWN on the applied basis, HOLD on the realized tape. The month’s cleanest finding: every covering house cut its applied multiple (Bernstein 45x to 40x, BofA 41x to 31x, Jefferies 40x to 35x) citing the AI-sector reset, yet PTs rose purely on the EPS raise - BofA held its PT unchanged as the pure cancel. The market’s realized forward P/E held roughly flat because EPS matched price, but EV/Sales stretched to the 93.6th percentile of 5.2 years. No multiple expansion is underwritten anywhere.
• Thesis anchor. The share-vs-Teradyne claim [6857-th-002] and its paired share claim [6857-th-003] moved from dormant to actively two-directional this month, but the eight post-print notes plus an ex-Advantest expert are un-extracted into evidence.yaml, so thesis_pulse still reads the share claim as 119 days stale. The valuation-risk claim [6857-th-005] (208 days stale) is the live bear case - price embeds ~16 years of value growth [news:nvda-cpo-spectrum-x-ramp].
• Highest-signal datapoint. The tester-TAM raise itself - Advantest led the entire semicap complex post-print (+42.5 percent vs peers +8-20 percent), the only positive name from Jun 30 to Aug 21 [evidence:6857:6857-ev-0220].
• Biggest gap. SoC book-to-bill [6857-kpi-003] has zero readings ever, the leading falsification input for the core demand claim - /thesis backfill 6857 (the evidence is already in corpus).

AMD - EPS up (out-year-led), multiple down on sentiment / hold on terminal

• EPS chain - revise UP. Revenue up, out-year-led: FY27 revenue +11.6 percent, FY27 EPS +14.6 percent, FY28 EPS +12.5 percent, with Data Center more than doubling YoY (fifth straight record server quarter); FY26 barely moved [evidence:AMD:amd-ev-0029] [evidence:AMD:amd-ev-0031]. Quality haircut: the +14pp YoY GAAP GM is partly an ~USD800mn MI308 base effect, the Q3 margin slope pauses flat at 56 percent and has been dodged four straight calls, and the shares/BVPS leg is deteriorating off-store (a vanished buyback line plus a USD4.75bn debt raise that closed 2026-08-17, still an unprocessed skeleton in Corpus) [evidence:AMD:amd-ev-0035].
• Multiple chain - revise DOWN on sentiment, HOLD on terminal. P/F-EPS -28.4 percent (60.9x to 43.6x), Fwd PEG -34.9 percent to 0.66x on a price down only ~10-14 percent - a textbook priced revision cycle, denominator-driven. Terminal held: the reverse-DCF still implies ~84 percent revenue CAGR vs a ~58 percent street.
• Thesis anchor. The valuation-risk claim [amd-th-005] still carries the stale pre-print 99.9th-percentile text; the de-rate has since pulled the actual percentile back to 82nd full-history. Both its deciding KPIs (slope, percentile) have zero readings despite being computable from feeds on hand - the binding gap.
• Highest-signal datapoint. An independent operator’s ~200k-800k/yr Instinct manufacturing ceiling [amd-ev-0035] - the first quantified supply cap on the design-win-conversion leg. The month’s real tape: MRVL +13 percent vs AMD -12 percent on a flat AI complex (divergence fired 2026-08-18, -11.3pp) - the market paid custom-ASIC over merchant #2.
• Biggest gap. Q2 FY26 print not merged into financials.csv and the USD4.75bn notes offering unprocessed - /process-filings AMD then /financials AMD.

MU - EPS hold, multiple hold

• EPS chain - HOLD. The June-print revision impulse has spent itself: FY27 estimates near-flat in-window (revenue +0.27 percent, EPS +1.03 percent) against the +22.6 percent FY27-EPS surge the prior month; FY28 still drifts up (EPS +3.09 percent) [slope:MU:EPS FY28 (USD)]. The margin leg holds at the record ~86 percent FQ4 guide, but management itself flagged the trajectory down - HBM has turned margin-dilutive and price is “harder and harder” to convert into margin [evidence:MU:mu-ev-0043] [evidence:MU:mu-ev-0045].
• Multiple chain - HOLD. Sentiment recovered off the mid-July trough (P/F-EPS +2.3 percent to 6.55x, price +5.0 percent outperforming a falling semis tape), but the terminal regime is unchanged - 6.55x forward sits in the 10.9th percentile of full history, a cyclical-trough multiple on peak earnings. The structural re-rating case gained a mechanism but no catalyst; management declined the ARR-style disclosure it needs [evidence:MU:mu-ev-0046].
• Thesis anchor. Thesis auto-promoted 2026-08-17, never analyst-adjudicated (all five claims confirmed_by_analyst null) - a live G-PIPE-88 item. The 2026-08-17 IR meeting forced two claim-language corrections (the SCA price/volume split and the HBM-margin reframe). The most counter-consensus disclosure of the month is HBM now being margin-dilutive [evidence:MU:mu-ev-0043].
• Highest-signal datapoint. HBM turning margin-dilutive [mu-ev-0043] - the disclosure that breaks the simple “HBM share equals margin” narrative.
• Biggest gap. No init suite exists for MU at all (no primer, debate, peers, or scans), and the FQ3 transcript is not in Corpus - /master-init MU is the real fix; near-term, ingest the transcript.

6981 Murata - EPS up, multiple down on sentiment / hold on terminal

• EPS chain - revise UP, for a lower-quality reason than the headline. Revenue confirmed and accelerating: FY27 guide raised to JPY2,110bn (+7.7 percent), the data-centre plan to +110 percent YoY, Q1 MLCC book-to-bill 1.47 [evidence:6981:6981-ev-0088]. But ~45 percent of the operating-profit raise is FX (JPY150 to 155); ex-FX the raise is only +5.9 percent, and mix uplift is being eaten by naphtha/raw-material cost inflation [evidence:6981:6981-ev-0089]. Shares/BVPS inert and unmeasured.
• Multiple chain - revise DOWN on sentiment, HOLD on terminal. Forward P/E de-rated 45.2x to 35.2x (-22.0 percent) while price fell only 9.9 percent - denominator again - yet it still sits in the 95.3rd percentile of five years. Peer co-movement: an MLCC-complex event, not AI (the cohort fell 38-54 percent while NVDA/TSMC/SPY were flat), and Murata was the persistent laggard (divergence fired 2026-08-13, -9.6pp). Terminal held: the reverse-DCF still solves a +21.8pp implied-vs-consensus revenue CAGR gap.
• Thesis anchor - the premise-vs-condition split. The revision-cycle claim [6981-th-002] has its premise broken on the near year: consensus FY27 EBIT (JPY450.2bn) sits 4.7 percent above management’s own JPY430bn guide, so the street is over-forecasting, not under-forecasting. But the claim’s falsification condition - “FY28 slope flat-to-down two quarters; margin below 27 percent” - is intact (FY28 EBIT +3.4 percent, margin above 30 percent). This must be logged as the distinction, not collapsed into one verdict. The mix-pricing claim [6981-th-004] moved toward falsification from three directions but is not breached (logged AMBIGUOUS, leaning challenge); the benign competing read is real and kept - softer price erosion is realised pricing power banked passively, and a TDK insider frames it as discretionary (“don’t want to make money for just one year”) [evidence:6981:6981-ev-0090].
• Highest-signal datapoint. A TDK senior sales director [6981-ev-0090] reporting industry book-to-bill above 2.0 and >50 percent allocation where 30-40 percent is normal, implying ~140 percent aggregate ordering, with the 2019 “dropped to zero” precedent - the first quantified instance of demand-KPI contamination on the name.
• Biggest gap. MLCC lead times [6981-kpi-003] and the secondary-market spot index [6981-kpi-011] both have zero readings, two of the four inputs that jointly falsify the reversibility claim - /deepen-corpus 6981 (China-channel check).

Shorts

ASML (short) - EPS up-modest, multiple hold

• EPS chain - revise UP, modestly - adverse to the short. In-window consensus rose on every line (FY27 EPS +0.99 percent, FY28 EPS +1.54 percent), but the 30-day revision slope flattened from ACCELERATING to STABLE on all but FY28 EPS [slope:ASML:EPS FY26 (EUR)] [expectations:ASML]. Level up, steepening stalled. The one new in-window document - a 2026-08-04 IR roadshow note - was never extracted into any store and runs adverse to the short on nearly every leg (a first-ever +30 percent/+30 percent two-year capacity guide with orders and down-payments banked, and a direct management rebuttal of the falling-litho-intensity TAM-shrink mechanism) [news:samsung-high-na-euv-deferral-to-1nm] [news:nanya-euv-adoption].
• Multiple chain - HOLD (range-bound). The big de-rate was a pre-window print-date denominator jump (P/F-EPS ~47x on 07-03 to 34x on the 07-15 print as guidance-raised EPS re-based). Within August the multiple V-shaped - trough 27.98x on 07-29, recovered to 31.66x by 08-19 - so ~a third of the collected de-rate reversed. valuation_scan_v4 codifies it as range-bound, mid-band.
• Thesis anchor. The peer tape is the real news and it is short-adverse: ASML fell least of the entire WFE complex (-12.5 percent peak-to-date vs KLAC -39, AMAT -32, LRCX -28), and the divergence engine fired it SHORT-adverse on 2026-07-29 (+13.9pp vs peers) - the complex sold off and the short’s name was the hedge everyone kept [news:asml-us-china-restriction-bill-stalled]. The 2026-08-02 standing review’s two prepared verbs (falsify the capacity claim [asml-th-006], log a held row for [asml-th-003]) are both unactioned.
• Highest-signal datapoint. ASML’s relative resilience through a -30-to-40 percent complex sell-off - the single most thesis-adverse fact of the month.
• Biggest gap. No EUV/DUV/High-NA gross-margin split in the canonical layer, the one input both the margin-stall short and the 2030 glidepath depend on - /extract-exposure ASML on the un-extracted 08-04 IR note, then /thesis backfill ASML.

CBRS (short) - EPS up, multiple down

• EPS chain - revise UP - adverse to the short. Revenue up: consensus stepped up on the print (CY26 +2.6 percent, CY27 +1.6 percent), and the FY26 core guide was raised to USD880-890mn from USD855-865mn - a second consecutive beat-and-raise [evidence:CBRS:cbrs-ev-0194] [evidence:CBRS:cbrs-ev-0204]. Margins up: FY26 core GM raised to 41-43 percent from 38-41 percent, with the consensus EBIT loss cut [evidence:CBRS:cbrs-ev-0207] [evidence:CBRS:cbrs-ev-0208]. Shares/BVPS dilutive - an OpenAI warrant partial exercise added 10.0mn Class N shares in July against a ~57 percent understated denominator.
• Multiple chain - revise DOWN. Sentiment sharply negative: the beat-and-raise was met with -25.2 percent since the print and a fresh post-print low. Peer co-movement idiosyncratic: excess -19.3pp vs the NVDA/AMD/AVGO basket. Terminal cut while numbers rose - Barclays 11x to 9x, Citi 15x to 14x - and the FY28 core-revenue range widened, dispersing rather than de-risking. Note the honest mechanical point: the -39 percent headline de-rate is mostly a -34pp NTM window-roll; the constant-basis compression is a milder -8.8 percent off the pre-print peak.
• Thesis anchor - the mis-specified test. The OpenAI-dependence claim [cbrs-th-005] is technically falsified (largest customer 34 percent, below the 50 percent line) but the test is mis-specified: the concentration fell only because the sovereign customer halved, while OpenAI itself went from 8.7 percent to 32 percent of revenue and 57 percent of receivables - substantively confirming [evidence:CBRS:cbrs-ev-0195]. The GM-disappointment claim [cbrs-th-002] is partially falsified on the hardware leg (core hardware GM 38.8 percent, above the challenge line). The lock-up-supply claim [cbrs-th-004] is confirming and moved away from falsification (USD197.9mn insider sales into the mid-August tranches). Recommend /thesis adjudicate to rewrite that claim’s condition to name OpenAI rather than “largest customer.”
• Highest-signal datapoint. The Q2 FY26 10-Q significant-customer and receivables tables - they adjudicate three claims at once (sovereign halved, re-concentration onto OpenAI, receivables up 145 percent against a -6.9 percent QoQ GAAP revenue decline) and not one of four covering banks engaged them.
• Biggest gap. No primary or channel evidence on the sovereign base (G42, MBZUAI), the mechanism at the heart of the revenue claim and half of H1 revenue - acquire via an expert call on the G42/MBZUAI relationship, then /extract-filings CBRS.

Brief-absent (26 of 34 held names)

No fresh brief this month, so no model call - each renders as a named gap, cleared by the nightly rotation (~4 names/night) or on demand. Run /research-deep --brief {T} to pull any name forward:

2330, 2454, 3533, LRCX, MKSI, 005930, ASMI, 0522, TER, ONTO, IFX, LITE, BE, ARM, CRDO, MRVL, SNVA, 2327, 6976, 8035, 6762, AVGO, 5536, CVLT, GIL, RBRK.

4 Adjudication and the acquisition agenda

The highest-probability read across the eight. August was a single regime expressed eight times: the numbers are still going up and the multiple is the thing that moved. Seven of eight EPS chains revised up, and in every de-rating name the compression was denominator-led - prices flat-to-down while estimates rose. What consensus is currently mispricing is not the direction of earnings; it is the durability of the multiple on names where the terminal assumption is doing the work. On the longs, the cleanest expression of that is the split the SPE/WFE stance now names: the revision slope has rotated toward test and process-control (6857 is the standout - a high-quality, FX-independent, TAM-led raise that led the entire complex), while broad-WFE and merchant-compute names (AMAT, AMD) are absorbing beats into a de-rate. 6857 is the highest-conviction long read of the month on evidence quality; MU is the clearest hold (the revision impulse is spent and the re-rating has a mechanism but no catalyst).

On the two shorts, the month ran adverse - and that is the finding, not a footnote. Both ASML and CBRS printed beat-and-raise, EPS revised up on both, and both saw the disconfirming datapoint land through a door the thesis did not name: ASML’s relative resilience through a 30-40 percent complex sell-off, and CBRS’s re-concentration onto OpenAI while the stated concentration test read as falsified. Neither short is dead - ASML’s multiple is range-bound and CBRS’s sentiment/peer/terminal legs all fell - but both theses are carrying a mis-specified or un-actioned condition that the analyst should resolve before the next print. This is exactly the AMBIGUOUS-in-the-challenge-direction discipline: the benign competing reads (ASML’s litho-intensity rebuttal, CBRS’s sovereign-driven optics) are real and are not argued away here.

The acquisition agenda, ranked by how much it would move a call:

1. Extract the two un-extracted, thesis-adverse documents now. ASML’s 2026-08-04 IR roadshow note and 6857’s eight post-print broker notes plus the ex-Advantest expert are in Corpus but invisible to every count-based reader - so thesis_pulse is reading both names as dormant while the corpus just contested them hard. /extract-exposure ASML, /thesis backfill 6857.
2. Re-tag KLAC’s evidence ledger to the post-2026-08-17 claim set - until then the KLAC pulse counts are structurally misleading and this review had to route around them. /thesis backfill KLAC.
3. Process AMD’s Q2 print and the USD4.75bn notes offering - the balance sheet the scan still calls “net cash” has changed. /process-filings AMD then /financials AMD.
4. Adjudicate the two premise-vs-condition splits before September grades them - rewrite the CBRS OpenAI-dependence claim [cbrs-th-005] to name OpenAI (/thesis adjudicate CBRS) and log the 6981 revision-cycle claim [6981-th-002] as premise-spent/condition-intact rather than a single verdict.
5. Fill the zero-reading deciding KPIs that make three claims unfalsifiable: AMD’s valuation slope/percentile (computable today), 6857’s SoC book-to-bill, and 6981’s MLCC lead-times and spot index.
6. The structural coverage gap: MU, a held long with the universe’s steepest revision impulse this year, has no init suite at all - /master-init MU.

For next month’s scorecard. These eight calls are now the baseline. The September edition grades them against the realized slope, stance and expectations moves - and by then the rotation will have lifted coverage well above eight, so the universe-level read stops being a bootstrap and starts being real.